Expand description
Deoptimisation guard.
When a GPU-offloaded kernel fails at runtime (CUDA error, numerical
divergence beyond an acceptable tolerance), the executor falls back to
the original CPU implementation. DeoptGuard tracks whether a given
blueprint is currently deopted so subsequent calls go straight to the
CPU path without retrying the GPU.
Structs§
- Deopt
Guard - Tracks deopted kernels by their blueprint fingerprint.
Enums§
- Deopt
Reason - Reason a particular kernel was deoptimised.
Type Aliases§
- Shared
Deopt Guard - Shared handle alias for the executor.